The Nvidia put explained: how one chipmaker became the AI economy's buyer of last resort

When analysts say "the Nvidia put," they mean this: if the AI buildout stumbles, Nvidia is the company most likely to step in and buy, lease, guarantee or invest. The phrase moved from trading desks to headlines after Jensen Huang resigned on November 12, 2026, because the person who set that policy is gone and nobody knows if his successor will keep it. This explainer defines the term, shows the real line items, and flags where the numbers disagree.

What a "put" means here

In finance, a put is a guaranteed buyer on the other side of a trade. Semafor columnist Liz Hoffman uses it for Nvidia's habit of stepping in as that buyer across the AI industry, writing that Huang was the sector's "central bank chairman," ready to back, acquire, hire, invest in or underwrite AI ventures. In her view the balance sheet absorbed market fears about AI and set a floor under deal prices. It kept AI financing and fundraising markets liquid, she says, though critics questioned the circular deals.

It is a metaphor, not a legal instrument. Nobody holds a contract that obliges Nvidia to rescue them. The "put" is the pattern of behavior, plus a set of real commitments that make it credible.

The pieces, according to the filings

CNBC's coverage of the resignation summarizes Nvidia's July filing. It showed $36 billion in off-balance-sheet backstops for neocloud customers and $20 billion tied to leases Nvidia takes on to rent out data center capacity. In August, CNBC reports, Nvidia offered a $105 billion backstop for OpenAI's rental of an Ohio campus; The Information reported that chief financial officer Colette Kress, now interim CEO, opposed that deal.

SemiAnalysis went through the same quarterly filing, which covers the period ended July 26, 2026. Its analysis, summarized by AI Weekly on September 11, lists the pieces separately:

• AI cloud take-or-pay floors of $36 billion.
• Data center leases Nvidia expects to reassign, about $20 billion.
• Guarantees of $108.5 billion for the Ohio campus, which AI Weekly describes as covering 4.25 gigawatts leased to OpenAI for 20 years.
• Supply and capacity commitments of $279 billion.
• A grand total of $530 billion of gross off-balance-sheet guarantees, up from $184 billion the prior quarter.

Those numbers use SemiAnalysis's definitions. The same AI Weekly summary says a floor price Nvidia guarantees on some cloud capacity is about $2.35 per GPU hour against a market rate of $4.50 to $4.60. That gap is the "put" in miniature: customers get a price floor and Nvidia keeps the upside if demand holds.

Why the numbers do not agree

Different sources get different totals because they count different things. CNBC cites analyst estimates of Nvidia's obligations to other firms at about $400 billion and notes some call it the "central bank of AI." Semafor says Nvidia guarantees more than $100 billion of data center leases, owns $51 billion in stock of private companies and $63 billion in listed ones. SemiAnalysis reaches $530 billion by summing six lines.

The key distinction is between money Nvidia pays only if something fails and contracts where it receives services in return. The $36 billion cloud agreements and $20 billion leases involve use of capacity or facilities, which is not the same as a guarantee that pays out on a default. SemiAnalysis itself frames the exposure as arising "only if the backstopped Neoclouds fail to meet their lease and offtake commitments." Add the lines together and you overstate any single risk, which is why we avoid giving one headline number.

The equity side

Beyond guarantees, CNBC lists Nvidia's investments: at least $30 billion in OpenAI, about $5 billion in Intel, and roughly $2 billion each in Nebius, Marvell, Lumentum, Coherent and CoreWeave. Nvidia also agreed to buy $10 billion of Anthropic stock right after Anthropic's IPO, which CNBC says is expected to start trading on November 19. When a supplier funds its own customers, critics call it circular financing. CNBC reports officials are monitoring that risk but see no imminent systemic danger.

Why the exit matters

Hoffman argues the successor faces much higher stakes and expects the board to be refreshed and to hold the new leader more closely accountable. Losing a buyer of last resort, she writes, could raise the cost of AI debt and equity and make the buildout more expensive for everyone. In October she wrote that Nvidia is "officially too big to fail" and asked what a stress test would show under higher AI debt costs, falling startup valuations and slower data center construction.

The market has begun to price that. Our report on the Nasdaq selloff tracked an earlier jolt, and CoreWeave's prepayment push shows neoclouds already asking customers to pay upfront. Federal Reserve Governor Michelle Bowman warned about AI debt in a speech we covered in our Bowman report.

Our take

The Nvidia put is real in the sense that matters: large, documented commitments that let weaker customers borrow against Nvidia's credit. It is overstated when a single number is used for it, because the lines mix guarantees, leases and ordinary supply contracts. The thing to watch is not the total but the direction. On November 17 Nvidia reports quarterly results; if the backstop and lease lines shrink or the language about customer support changes, the put is being withdrawn. If they grow under an interim CEO, the new management is choosing to keep it.